siddharthmb/2026.MI.Sibling-Experiments

The MATS-II sibling experiments of June and July 2026: outputs, logs and small artifacts

Everything at the top of the MATS-II scratch tree that is not the rational-agents programme and not weights, one tarball per directory: the conception and steering work (conv_steering, conv_dataset, conv_MoE, conv_test, partner_identity, reasoning_convergence, interlens_refactor), the multi-model failure-prediction and back-prop communication experiments (multimodel-predict-failure-with-interp, multimodel-backprop, mpf-adv), the acausal-coordination battery, the oversight gradient attack, the judge-debate verification, the B200 session's results, products and sibling logs (b200_final, b200_scaled, b200_sibling_logs), the chat-harness validations, the figure-fix build logs, the scifact data copy, and the loose top-level files (_toplevel_files.tar.zst: the light judge sweep script, two download logs, the two QKV state notes and a train-to-steer elicitation summary written as an email). The adapters these runs trained are in 2026.MI.RL-Steering-Adapters and 2026.MI.Ladder-Target-Gating; the transcripts in 2026.MI.Transcripts.

Families

• experiments: one tarball per sibling experiment directory of the MATS-II scratch tree (June and July 2026)
• toplevel_files: the loose files at the top of ii_mats/: the light judge sweep script, two download logs, the two QKV state notes, a train-to-steer summary written as an email

Contents

21 tarballs, 5985 files, 2.51 GB unpacked, 2.12 GB here. Each tarball unpacks to the directory it is named after (member paths start with that name); MANIFEST.json lists every tarball with its sha256, file count, sizes, newest modification time and, for code snapshots, the git HEAD the tree was at.

family	tarball	source entry	files	raw GB	packed GB	newest file
experiments	experiments/acausal_coordination.tar.zst	acausal_coordination	34	0.07	0.00	2026-08-24
experiments	experiments/b200_final.tar.zst	b200_final	102	0.00	0.00	2026-06-26
experiments	experiments/b200_scaled.tar.zst	b200_scaled	95	0.00	0.00	2026-06-26
experiments	experiments/b200_sibling_logs.tar.zst	b200_sibling_logs	25	0.00	0.00	2026-06-28
experiments	experiments/chat_harness.tar.zst	chat_harness	156	0.00	0.00	2026-07-01
experiments	experiments/conv_MoE.tar.zst	conv_MoE	122	0.01	0.01	2026-07-15
experiments	experiments/conv_dataset.tar.zst	conv_dataset	672	0.24	0.03	2026-08-17
experiments	experiments/conv_steering.tar.zst	conv_steering	1491	0.89	0.86	2026-07-06
experiments	experiments/conv_test.tar.zst	conv_test	3	0.00	0.00	2026-07-02
experiments	experiments/data.tar.zst	data	18	0.01	0.01	2026-07-15
experiments	experiments/figfix_umb.tar.zst	figfix_umb	78	0.00	0.00	2026-09-03
experiments	experiments/interlens_refactor.tar.zst	interlens_refactor	80	0.00	0.00	2026-07-06
experiments	experiments/judge_debate_verify.tar.zst	judge_debate_verify	5	0.00	0.00	2026-06-24
experiments	experiments/mpf-adv.tar.zst	mpf-adv	2134	0.53	0.49	2026-07-16
experiments	experiments/multimodel-backprop.tar.zst	multimodel-backprop	38	0.00	0.00	2026-07-06
experiments	experiments/multimodel-predict-failure-with-interp.tar.zst	multimodel-predict-failure-with-interp	709	0.46	0.44	2026-07-06
experiments	experiments/negotiation_bridge.tar.zst	negotiation_bridge	2	0.00	0.00	2026-06-24
experiments	experiments/oversight_gradient_attack.tar.zst	oversight_gradient_attack	6	0.00	0.00	2026-07-15
experiments	experiments/partner_identity.tar.zst	partner_identity	4	0.27	0.27	2026-07-09
experiments	experiments/reasoning_convergence.tar.zst	reasoning_convergence	207	0.01	0.00	2026-07-13
toplevel_files	toplevel_files/_toplevel_files.tar.zst	./(top-level files)	4	0.00	0.00	2026-06-25

How to unpack

hf download siddharthmb/2026.MI.Sibling-Experiments --repo-type dataset --local-dir 2026.MI.Sibling-Experiments
cd 2026.MI.Sibling-Experiments && for f in */*.tar.zst; do tar -I zstd -xf "$f"; done   # or one tarball at a time

zstd 1.4 or newer; GNU tar. A tarball's members are relative paths starting with the source entry's name, so unpacking several in one directory recreates the workspace layout.

Related repositories

• The collection: https://huggingface.co/collections/siddharthmb/2026-mats-rational-agents-6acc1dd5623e8e00176b80a8
• The workspace index: https://huggingface.co/datasets/siddharthmb/2026.RA.Workspace-Index

Provenance

Archived from /nlp/scr/siddharth/ii_mats on 2026-10-11 by the MATS archive job 17819018 of the transcoder-adapters cleanup (report claude_updates/memory/reports/mats_archive.md in that repository). Run directories were packed unchanged (tar --sort=name | zstd -10) except that .git/ directories (histories live in the private mats-ii repository), .cache/ upload bookkeeping and __pycache__/ were left out of every tarball. Every file name and every text file was scanned for credentials before upload (Hub, Anthropic, OpenAI, AWS and W&B key shapes, private-key headers, credential URLs, paths of the secrets store); the report lists what was excluded. The local copy was deleted after every uploaded file was verified on the Hub by name, size and hash, and a <name>.txt breadcrumb naming this repository was left beside it.